Skip to content

Nandi

Agent Hub

AI agents and packs, ready to install in Nandi. Browse them here, then install from Library › Browse in the app. Each one declares the tools, hosts, and OS control it needs, and you approve them before it first runs. Connectors link agents to the services you use.

Browse the catalog

    Connectors

    A connector links your agents to a service, such as an issue tracker or a web page reader. Each one says what it can read and write, the hosts it reaches, and the sign-in it needs. Installing a connector gives no agent access. An agent gets access when you allow it, once or for that agent, with read and write kept separate. You can remove access in Settings › Connectors › Access.

      Publish Your Own Agent

      An agent is a folder with a skill.md in a public GitHub repo. Publishing adds an entry for it to the registry, which is itself a git repo: you open a pull request, and maintainers review what the agent asks for before merging. To publish a connector, see Publish a Connector.

      Write skill.md

      The front matter declares who the agent is and everything it may use. The body is its system prompt.

      ---
      name: Invoice Reader
      description: Reads PDF invoices in your working folder and saves their totals to a CSV.
      allowed-tools: [shell, read_file, write_file, list_dir]
      requires_approval: [write_file]
      network:
        allow:
          - host: api.exchangerate.host
      ---
      
      You are Invoice Reader. You extract the vendor, date, currency and total
      from each PDF in ./invoices/ and append one row per invoice to
      ./invoices/totals.csv.
      
      - Never delete or modify the PDFs.
      - If a total is unreadable, list the file and say why instead of guessing.
      - Only api.exchangerate.host is reachable. If the user asks for another
        source, explain that they can fork this agent and add the host to
        network.allow.
      • name and description appear in the Agent Hub, in the app's Library, and in the registry entry.
      • allowed-tools lists every tool the agent may call. Anything else is refused.
      • requires_approval pauses the run for your confirmation before these tools act.
      • network.allow lists the hosts it may reach through the egress proxy. Keep it as short as possible.
      • The body after the front matter is the system prompt. Say what the agent does, what it must never do, and what to tell the user when it can't.

      Test it locally

      Put the folder in your agents repo (~/.nandi/repo/.nandi/agents/<id>/) and Nandi picks it up right away. Chat with it, check what it asks you to approve, and confirm its prerequisites from a terminal:

      nandi verify --agent invoice-reader

      To test a registry entry before you open the pull request, add your fork of the registry under Settings › Sources and install from it in Library › Browse.

      Tag a release

      Push the folder to a public GitHub repo and tag the release. A repo that holds one agent uses v<version>; a repo with several uses <id>-v<version>.

      git tag v1.0.0        # or invoice-reader-v1.0.0 in a repo with several agents
      git push origin v1.0.0

      Add it to the registry

      1. Fork the registry repo.
      2. Add agents/<id>.json matching the agent schema (example below).
      3. If you're new, add publishers/<your-handle>.json.
      4. Add the same entry to the agents list in index.json.
      5. Open a pull request. Maintainers review the capabilities before merging.

      agents/invoice-reader.json

      {
        "id": "invoice-reader",
        "name": "Invoice Reader",
        "description": "Reads PDF invoices in your working folder and saves their totals to a CSV.",
        "publisher": "your-handle",
        "repo": "github.com/your-handle/nandi-agents",
        "path": "agents/invoice-reader",
        "latest_version": "1.0.0",
        "tags": ["finance", "pdf"],
        "min_nandi_version": "0.1.0",
        "capabilities": {
          "tools": ["shell", "read_file", "write_file", "list_dir"],
          "network": ["api.exchangerate.host"],
          "pilots": []
        },
        "icon_url": "https://raw.githubusercontent.com/your-handle/nandi-agents/HEAD/agents/invoice-reader/icon.svg",
        "readme_url": "https://raw.githubusercontent.com/your-handle/nandi-agents/HEAD/agents/invoice-reader/README.md"
      }

      publishers/your-handle.json

      {
        "github_handle": "your-handle",
        "display_name": "Your Name",
        "public_key": "ed25519:<printed by nandi keygen>",
        "created_at": "2026-10-06T00:00:00Z"
      }

      The full field list, including optional and reserved fields, is in the registry schema.

      Shortcut: Publish, in the agent's ⋯ menu in Nandi, does all of this for you. It tags the release, builds the entry from skill.md, signs it if you have a key, and opens the pull request.

      Open the registry repo

      Sign it (recommended)

      Signing is optional but recommended. Your public key goes in publishers/<handle>.json; each entry carries the commit and the signature over it. Signed entries show as verified, and Nandi refuses an install if the tag has moved since you signed it.

      nandi keygen     # once: saves ~/.nandi/keys/publisher.key, prints the public key
      nandi sign --kind agent --id invoice-reader --release 1.0.0 \
                  --repo github.com/your-handle/nandi-agents
      # paste the printed "commit" and "signature" into agents/invoice-reader.json

      Best practices

      • Declare the least you need

        Users approve every tool, host, and pilot capability. A short list gets approved; a long one gets uninstalled. Never list a host you don't call.

      • Ask before acting

        Put anything that writes, sends, deletes, or spends in requires_approval. Reading is usually fine without it.

      • Name things plainly

        IDs are lowercase letters, digits, and hyphens, and become folder names, so make them describe the job (invoice-reader, not my-agent-2).

      • Sign your releases

        Create a key once with nandi keygen, then sign each release with nandi sign.

      • Ship a README

        Say what the agent does, what it needs (files, hosts, keys), and a sample prompt. The README is shown in Nandi.

      Publish a Connector

      Anyone can publish a connector. A connector is one connector.md file in a public GitHub repo: the MCP server it starts, what it can reach, and plain instructions for agents. You publish it the same way as an agent, with a pull request to the registry that maintainers review.

      Write connector.md

      The front matter says how to start the server and everything it can touch. The body tells agents how to use it.

      ---
      name: Issue tracker
      description: Read and create issues in your issue tracker.
      version: 1.2.0
      run: npx @acme/tracker-mcp@1.2.0
      auth: api_key
      hosts:
        - api.tracker.example.com
      tools:
        - name: search_issues
          description: Search issues by text
          access: read
        - name: create_issue
          description: Create an issue
          access: write
      reads:
        - Your issues and projects
      writes:
        - New issues
      tags: [issues, project-management]
      ---
      
      Search before you create an issue, so you don't add a duplicate.
      Say which project a new issue goes to. If you are not sure, ask.
      • run starts the server from an exact version: npx <package>@<version> or uvx <package>==<version>. A remote server uses url (https only) in its place.
      • auth is none, api_key, oauth or device_code. Keys stay in the OS keyring. List OAuth scopes when it asks for any.
      • hosts lists every host the server reaches. The container reaches only these.
      • tools lists every tool, each marked read or write. A tool you leave out counts as write.
      • reads and writes say in plain words what it can see and change. They are shown when someone installs it.
      • image is only for a server that needs something unusual, such as a browser. placement is container, the only one the public registry accepts.

      Push it and tag a release

      Push the file to a public GitHub repo and tag the release. A repo with one connector uses v<version>; a repo with several uses <id>-v<version>.

      git tag issue-tracker-v1.2.0
      git push origin issue-tracker-v1.2.0

      Sign it

      Sign the release with your publisher key. Signed connectors show as verified, and Nandi refuses an install if the tag has moved since you signed it.

      nandi sign --kind connector --id issue-tracker --release 1.2.0 \
                 --commit <sha> --key ~/.nandi/keys/publisher.key
      # paste the printed "commit" and "signature" into connectors/issue-tracker.json

      Open the pull request

      1. Fork the registry repo.
      2. Add connectors/<id>.json built from your connector.md, with its commit and signature.
      3. If you're new, add publishers/<your-handle>.json with your public key.
      4. Add the same entry to the connectors list in index.json.
      5. Open a pull request. Maintainers review it against the rules below.

      The full field list is in the registry schema.

      Shortcut: Studio › Publish in Nandi does all of this for you. It builds connector.md from an MCP server you already use, tags and signs the release, and opens the pull request.

      Open the registry repo

      What reviewers check

      • An exact version

        The package is pinned to one version, such as 1.2.0. No ranges, tags or latest.

      • Every host listed

        Each host the server calls is in hosts, and no host it doesn't call.

      • Honest tool access

        A tool that sends, creates, changes or deletes anything is marked write. Only tools that just look are read.

      • Container only

        Public connectors run in a container. A connector that runs on your computer is added by hand or from a company registry.